Dialectic
A sidecar that checks an agent's run at three gates: did the sub-goals cover the goal, did each action serve a sub-goal, and did the result match what was expected. When a gate fails, it suggests the smallest change to the plan that would fix it.
How it works
The agent declares a goal and an expected result. Dialectic reads the run trace and checks three gates. Decomposition: every key term of the goal appears in at least one sub-goal. Planning: every action is tied to a sub-goal, and every sub-goal has an action. Outcome: the observed result has the fields and values the agent said to expect. Each failure comes with a suggested fix. This demo uses plain rules; the proposal's "Counterfactual Synthesis Engine" would need an LLM judge to do this on free text.
Format: GOAL, EXPECT key=value, SUB s1: …, ACT s1: …, RESULT key=value
Findings and suggested fixes
Claimed by the proposal, not verified
- "40% decrease in post-deployment behavioral revisions" from "early adopters". No adopters, baseline or study are named.
- "Reductions in regulatory compliance issues." No figures or setting given.
- It "intercepts low-level predictions" during inference. Hosted models like Gemini, ChatGPT and Claude don't expose internals to a sidecar. In practice it can only read prompts, tool calls and outputs, as this demo does.
- A "running probability distribution over plausible intent mappings" is described but not defined.
What to try
- Load the expense summary trace. The decomposition gate flags that "weekly" never made it into a sub-goal, and the outcome gate catches the wrong date range.
- Load the key rotation trace. It passes decomposition and planning but the result shows the old key still active.
- Edit the morning package trace: delete the line "ACT s3" and check again. The planning gate flags s3 as having no action.
- Change a RESULT value to match its EXPECT line to see a gate turn green.
What a real build would need
Agents that log goal, plan and expected result in a fixed format before acting (most of the value is here, and it's cheap), an LLM judge for free-text checks, and a place to send findings, such as the Port 8899 hub or a weekly digest to the board.